Frequency-aware Adaptive Contrastive Learning for Sequential Recommendation

TL;DR

This work examines contrastive learning for sequential recommendation and reveals a bias against low-frequency items introduced by standard data augmentation. It introduces FACL, which combines micro-level adaptive perturbations that protect rare items with macro-level reweighting that emphasizes sparse and short sequences during training. Empirical results across six public datasets show that FACL consistently outperforms both data-augmentation and model-augmentation baselines, achieving up to 3–4% improvements and stronger robustness to long-tail distributions. The approach provides a practical path to preserving intent signals in real-world, long-tail recommendation scenarios, with broad applicability to sparse-user settings.

Abstract

In this paper, we revisited the role of data augmentation in contrastive learning for sequential recommendation, revealing its inherent bias against low-frequency items and sparse user behaviors. To address this limitation, we proposed FACL, a frequency-aware adaptive contrastive learning framework that introduces micro-level adaptive perturbation to protect the integrity of rare items, as well as macro-level reweighting to amplify the influence of sparse and rare-interaction sequences during training. Comprehensive experiments on five public benchmark datasets demonstrated that FACL consistently outperforms state-of-the-art data augmentation and model augmentation-based methods, achieving up to 3.8% improvement in recommendation accuracy. Moreover, fine-grained analyses confirm that FACL significantly alleviates the performance drop on low-frequency items and users, highlighting its robust intent-preserving ability and its superior applicability to real-world, long-tail recommendation scenarios.

Figures (8)

• Figure 1: Left: Proportion of items completely dropped under drop-item perturbation, across item frequency bins (baseline vs. ours). Right: Prediction accuracy of target items at different frequency levels (baseline vs. ours).
• Figure 2: Overview of the proposed FACL framework.
• Figure 4: Parameter sensitivity analysis for (a) $\gamma$, (b) $\eta$, and (c) $\beta$ on ML-20M and Yelp.
• Figure 5: HR@10 grouped by item frequency (low, medium, high) on ML-20M (left) and Yelp (right).
• Figure 6: HR@10 grouped by user frequency (low, medium, high) on ML-20M (left) and Yelp (right).